In September 2024 Microsoft announced a 20-year power deal, starting in 2028, with Constellation Energy. By then, the energy demands of AI had become clear.
The deal was to get Unit 1 of the old Three Mile Island nuclear reactor back up and running after it was shut down in 2019. The plan was to use the clean nuclear power for Microsoft’s data centres, which would be used by OpenAI’s ChatGPT.
Energy has been a big hurdle for new data centres; several have been built and are ready, but remain unable to operate due to lack of power.
As a result, the share prices of energy stocks have surged, with Constellation more than doubling since the 2024 announcement. This falls under the theme of AI-allied trades, where memory chip and hard drive manufacturers have been the big winners for investors.
But while chip prices have rocketed, the energy story is starting to unravel a bit.
The first area of concern is that there may have been some double-counting. Here’s an example: five proposed new data centres in one US state reportedly needed 4GW of power; two got approved, and the power demand was less than 2GW.
Add to this list the dominant players with the cash to roll out their AI models: Microsoft and Alphabet
Second, Nvidia unveiled its newest chip, the Nvidia Rubin, at the Consumer Electronics Show in Las Vegas this month. This chip is vastly more power-efficient than the H200 chip now being used. Exact data has not been released, and the chip is not yet commercially available, so real-world testing is not possible, but Nvidia is positioning it as a higher performance-per-watt chip. In other words, more bang for buck.
Third, research firm CreditSights recently published a report suggesting that US energy providers are adding more capacity than what is actually needed for new AI data centres — and this is before the introduction of the new, more energy-efficient chips.
The lesson here is that while going for AI-allied trades such as energy is a clever move, it still comes with challenges, risks and bubbles. In many ways, it suggests that the core AI investment theme should perhaps still be focused on the large players that offer what nobody else has.
They include Nvidia, with its chips that are streets ahead of everybody else’s; the Taiwan Semiconductor Manufacturing Co (TSMC), the world’s best manufacturer of chips for Nvidia; and ASML, which has the extreme ultraviolet lithography tech that is used by TSMC in chip production.
Add to this list the dominant players with the cash to roll out their AI models: Microsoft and Alphabet.
So, it’s back to keeping it simple, given that it seems the allied plays have all run really hard and there is a real risk that demand may not be as high as expected.
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.